void cv::gpu::reduce(const GpuMat& src, GpuMat& dst, int dim, int reduceOp, int dtype, Stream& stream)
{
using namespace ::cv::gpu::device::matrix_reductions;
CV_Assert(src.depth() <= CV_32F && src.channels() <= 4 && dtype <= CV_32F);
CV_Assert(dim == 0 || dim == 1);
CV_Assert(reduceOp == CV_REDUCE_SUM || reduceOp == CV_REDUCE_AVG || reduceOp == CV_REDUCE_MAX || reduceOp == CV_REDUCE_MIN);
if (dtype < 0)
dtype = src.depth();
dst.create(1, dim == 0 ? src.cols : src.rows, CV_MAKETYPE(dtype, src.channels()));
if (dim == 0)
{
typedef void (*caller_t)(const PtrStepSzb& src, const PtrStepSzb& dst, int reduceOp, cudaStream_t stream);
static const caller_t callers[6][6] =
{
{
reduceRows_gpu<unsigned char, int, unsigned char>,
0/*reduceRows_gpu<unsigned char, int, signed char>*/,
0/*reduceRows_gpu<unsigned char, int, unsigned short>*/,
0/*reduceRows_gpu<unsigned char, int, short>*/,
reduceRows_gpu<unsigned char, int, int>,
reduceRows_gpu<unsigned char, int, float>
},
{
0/*reduceRows_gpu<signed char, int, unsigned char>*/,
0/*reduceRows_gpu<signed char, int, signed char>*/,
0/*reduceRows_gpu<signed char, int, unsigned short>*/,
0/*reduceRows_gpu<signed char, int, short>*/,
0/*reduceRows_gpu<signed char, int, int>*/,
0/*reduceRows_gpu<signed char, int, float>*/
},
{
0/*reduceRows_gpu<unsigned short, int, unsigned char>*/,
0/*reduceRows_gpu<unsigned short, int, signed char>*/,
reduceRows_gpu<unsigned short, int, unsigned short>,
0/*reduceRows_gpu<unsigned short, int, short>*/,
reduceRows_gpu<unsigned short, int, int>,
reduceRows_gpu<unsigned short, int, float>
},
{
0/*reduceRows_gpu<short, int, unsigned char>*/,
0/*reduceRows_gpu<short, int, signed char>*/,
0/*reduceRows_gpu<short, int, unsigned short>*/,
reduceRows_gpu<short, int, short>,
reduceRows_gpu<short, int, int>,
reduceRows_gpu<short, int, float>
},
{
0/*reduceRows_gpu<int, int, unsigned char>*/,
0/*reduceRows_gpu<int, int, signed char>*/,
0/*reduceRows_gpu<int, int, unsigned short>*/,
0/*reduceRows_gpu<int, int, short>*/,
reduceRows_gpu<int, int, int>,
reduceRows_gpu<int, int, float>
},
{
0/*reduceRows_gpu<float, float, unsigned char>*/,
0/*reduceRows_gpu<float, float, signed char>*/,
0/*reduceRows_gpu<float, float, unsigned short>*/,
0/*reduceRows_gpu<float, float, short>*/,
0/*reduceRows_gpu<float, float, int>*/,
reduceRows_gpu<float, float, float>
}
};
const caller_t func = callers[src.depth()][dst.depth()];
if (!func)
CV_Error(CV_StsUnsupportedFormat, "Unsupported combination of input and output array formats");
func(src.reshape(1), dst.reshape(1), reduceOp, StreamAccessor::getStream(stream));
}
else
{
typedef void (*caller_t)(const PtrStepSzb& src, int cn, const PtrStepSzb& dst, int reduceOp, cudaStream_t stream);
static const caller_t callers[6][6] =
{
{
reduceCols_gpu<unsigned char, int, unsigned char>,
0/*reduceCols_gpu<unsigned char, int, signed char>*/,
0/*reduceCols_gpu<unsigned char, int, unsigned short>*/,
0/*reduceCols_gpu<unsigned char, int, short>*/,
reduceCols_gpu<unsigned char, int, int>,
reduceCols_gpu<unsigned char, int, float>
},
{
0/*reduceCols_gpu<signed char, int, unsigned char>*/,
0/*reduceCols_gpu<signed char, int, signed char>*/,
0/*reduceCols_gpu<signed char, int, unsigned short>*/,
0/*reduceCols_gpu<signed char, int, short>*/,
0/*reduceCols_gpu<signed char, int, int>*/,
0/*reduceCols_gpu<signed char, int, float>*/
},
{
0/*reduceCols_gpu<unsigned short, int, unsigned char>*/,
0/*reduceCols_gpu<unsigned short, int, signed char>*/,
reduceCols_gpu<unsigned short, int, unsigned short>,
0/*reduceCols_gpu<unsigned short, int, short>*/,
reduceCols_gpu<unsigned short, int, int>,
reduceCols_gpu<unsigned short, int, float>
},
{
0/*reduceCols_gpu<short, int, unsigned char>*/,
0/*reduceCols_gpu<short, int, signed char>*/,
0/*reduceCols_gpu<short, int, unsigned short>*/,
reduceCols_gpu<short, int, short>,
reduceCols_gpu<short, int, int>,
reduceCols_gpu<short, int, float>
},
{
0/*reduceCols_gpu<int, int, unsigned char>*/,
0/*reduceCols_gpu<int, int, signed char>*/,
0/*reduceCols_gpu<int, int, unsigned short>*/,
0/*reduceCols_gpu<int, int, short>*/,
reduceCols_gpu<int, int, int>,
reduceCols_gpu<int, int, float>
},
{
0/*reduceCols_gpu<float, unsigned char>*/,
0/*reduceCols_gpu<float, signed char>*/,
0/*reduceCols_gpu<float, unsigned short>*/,
0/*reduceCols_gpu<float, short>*/,
0/*reduceCols_gpu<float, int>*/,
reduceCols_gpu<float, float, float>
}
};
const caller_t func = callers[src.depth()][dst.depth()];
if (!func)
CV_Error(CV_StsUnsupportedFormat, "Unsupported combination of input and output array formats");
func(src, src.channels(), dst, reduceOp, StreamAccessor::getStream(stream));
}
}
void cv::gpu::reduce(const GpuMat& src, GpuMat& dst, int dim, int reduceOp, int dtype, Stream& stream)
{
CV_Assert( src.channels() <= 4 );
CV_Assert( dim == 0 || dim == 1 );
CV_Assert( reduceOp == CV_REDUCE_SUM || reduceOp == CV_REDUCE_AVG || reduceOp == CV_REDUCE_MAX || reduceOp == CV_REDUCE_MIN );
if (dtype < 0)
dtype = src.depth();
dst.create(1, dim == 0 ? src.cols : src.rows, CV_MAKE_TYPE(CV_MAT_DEPTH(dtype), src.channels()));
if (dim == 0)
{
typedef void (*func_t)(PtrStepSzb src, void* dst, int op, cudaStream_t stream);
static const func_t funcs[7][7] =
{
{
::reduce::rows<unsigned char, int, unsigned char>,
0/*::reduce::rows<unsigned char, int, signed char>*/,
0/*::reduce::rows<unsigned char, int, unsigned short>*/,
0/*::reduce::rows<unsigned char, int, short>*/,
::reduce::rows<unsigned char, int, int>,
::reduce::rows<unsigned char, float, float>,
::reduce::rows<unsigned char, double, double>
},
{
0/*::reduce::rows<signed char, int, unsigned char>*/,
0/*::reduce::rows<signed char, int, signed char>*/,
0/*::reduce::rows<signed char, int, unsigned short>*/,
0/*::reduce::rows<signed char, int, short>*/,
0/*::reduce::rows<signed char, int, int>*/,
0/*::reduce::rows<signed char, float, float>*/,
0/*::reduce::rows<signed char, double, double>*/
},
{
0/*::reduce::rows<unsigned short, int, unsigned char>*/,
0/*::reduce::rows<unsigned short, int, signed char>*/,
::reduce::rows<unsigned short, int, unsigned short>,
0/*::reduce::rows<unsigned short, int, short>*/,
::reduce::rows<unsigned short, int, int>,
::reduce::rows<unsigned short, float, float>,
::reduce::rows<unsigned short, double, double>
},
{
0/*::reduce::rows<short, int, unsigned char>*/,
0/*::reduce::rows<short, int, signed char>*/,
0/*::reduce::rows<short, int, unsigned short>*/,
::reduce::rows<short, int, short>,
::reduce::rows<short, int, int>,
::reduce::rows<short, float, float>,
::reduce::rows<short, double, double>
},
{
0/*::reduce::rows<int, int, unsigned char>*/,
0/*::reduce::rows<int, int, signed char>*/,
0/*::reduce::rows<int, int, unsigned short>*/,
0/*::reduce::rows<int, int, short>*/,
::reduce::rows<int, int, int>,
::reduce::rows<int, float, float>,
::reduce::rows<int, double, double>
},
{
0/*::reduce::rows<float, float, unsigned char>*/,
0/*::reduce::rows<float, float, signed char>*/,
0/*::reduce::rows<float, float, unsigned short>*/,
0/*::reduce::rows<float, float, short>*/,
0/*::reduce::rows<float, float, int>*/,
::reduce::rows<float, float, float>,
::reduce::rows<float, double, double>
},
{
0/*::reduce::rows<double, double, unsigned char>*/,
0/*::reduce::rows<double, double, signed char>*/,
0/*::reduce::rows<double, double, unsigned short>*/,
0/*::reduce::rows<double, double, short>*/,
0/*::reduce::rows<double, double, int>*/,
0/*::reduce::rows<double, double, float>*/,
::reduce::rows<double, double, double>
}
};
const func_t func = funcs[src.depth()][dst.depth()];
if (!func)
CV_Error(CV_StsUnsupportedFormat, "Unsupported combination of input and output array formats");
func(src.reshape(1), dst.data, reduceOp, StreamAccessor::getStream(stream));
}
else
{
typedef void (*func_t)(PtrStepSzb src, void* dst, int cn, int op, cudaStream_t stream);
static const func_t funcs[7][7] =
{
{
::reduce::cols<unsigned char, int, unsigned char>,
0/*::reduce::cols<unsigned char, int, signed char>*/,
0/*::reduce::cols<unsigned char, int, unsigned short>*/,
0/*::reduce::cols<unsigned char, int, short>*/,
::reduce::cols<unsigned char, int, int>,
::reduce::cols<unsigned char, float, float>,
::reduce::cols<unsigned char, double, double>
},
{
0/*::reduce::cols<signed char, int, unsigned char>*/,
0/*::reduce::cols<signed char, int, signed char>*/,
0/*::reduce::cols<signed char, int, unsigned short>*/,
0/*::reduce::cols<signed char, int, short>*/,
0/*::reduce::cols<signed char, int, int>*/,
0/*::reduce::cols<signed char, float, float>*/,
0/*::reduce::cols<signed char, double, double>*/
},
{
0/*::reduce::cols<unsigned short, int, unsigned char>*/,
0/*::reduce::cols<unsigned short, int, signed char>*/,
::reduce::cols<unsigned short, int, unsigned short>,
0/*::reduce::cols<unsigned short, int, short>*/,
::reduce::cols<unsigned short, int, int>,
::reduce::cols<unsigned short, float, float>,
::reduce::cols<unsigned short, double, double>
},
{
0/*::reduce::cols<short, int, unsigned char>*/,
0/*::reduce::cols<short, int, signed char>*/,
0/*::reduce::cols<short, int, unsigned short>*/,
::reduce::cols<short, int, short>,
::reduce::cols<short, int, int>,
::reduce::cols<short, float, float>,
::reduce::cols<short, double, double>
},
{
0/*::reduce::cols<int, int, unsigned char>*/,
0/*::reduce::cols<int, int, signed char>*/,
0/*::reduce::cols<int, int, unsigned short>*/,
0/*::reduce::cols<int, int, short>*/,
::reduce::cols<int, int, int>,
::reduce::cols<int, float, float>,
::reduce::cols<int, double, double>
},
{
0/*::reduce::cols<float, float, unsigned char>*/,
0/*::reduce::cols<float, float, signed char>*/,
0/*::reduce::cols<float, float, unsigned short>*/,
0/*::reduce::cols<float, float, short>*/,
0/*::reduce::cols<float, float, int>*/,
::reduce::cols<float, float, float>,
::reduce::cols<float, double, double>
},
{
0/*::reduce::cols<double, double, unsigned char>*/,
0/*::reduce::cols<double, double, signed char>*/,
0/*::reduce::cols<double, double, unsigned short>*/,
0/*::reduce::cols<double, double, short>*/,
0/*::reduce::cols<double, double, int>*/,
0/*::reduce::cols<double, double, float>*/,
::reduce::cols<double, double, double>
}
};
const func_t func = funcs[src.depth()][dst.depth()];
if (!func)
CV_Error(CV_StsUnsupportedFormat, "Unsupported combination of input and output array formats");
func(src, dst.data, src.channels(), reduceOp, StreamAccessor::getStream(stream));
}
}